The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Skillmatch MCP listing page.
Claude-powered job fit analyzer for people who prove their skills through work, not credentials.
SkillMatch is an MCP server that gives Claude access to your GitHub portfolio, resume, and job search preferences. It lets Claude:
All data stays local on your machine. Nothing is sent anywhere except public GitHub API calls to fetch your own repos.
Speaks MCP protocol revision 2026-07-28 over stdio, on SDK mcp>=2.0,<3. It is stateless:
no initialize handshake and no session id, with every request carrying its own protocol version
and client identity. tools/list and server/discover return ttlMs/cacheScope freshness
hints, and tool order is stable across calls so client-side and prompt caches keep hitting.
Upgrading from 1.x: 2.0.0 drops the
initializehandshake, so it requires a client that speaks2026-07-28. Tool names, schemas, and results are unchanged — only the transport moved. Stay on 1.x if you need to serve an older client.
Claude Code (CLI):
Then restart Claude Code.
Claude Desktop:
Open your Claude Desktop configuration file and add SkillMatch to the mcpServers section:
Replace /path/to/skillmatch-mcp/server.py with the actual path to server.py on your machine, then restart Claude Desktop.
After restarting, SkillMatch tools will be available in your conversation. Works in Claude Code and Claude Desktop.
The first time you use SkillMatch, Claude will walk you through a setup conversation. It looks something like this:
After onboarding, your profile is saved locally and reused in every future conversation.
You can enrich your profile at any time by asking Claude to update it. These fields help Claude reason about nuanced fit:
| Field | Type | Example |
|---|---|---|
work_style | object | {async_preferred: true, ic_vs_leadership: "ic", client_facing_tolerance: "none", team_size_preference: "small"} |
optimizing_for | array | ["comp", "autonomy", "interesting_problems"] |
unlisted_skills | array | ["MCP protocol implementation", "Windows installer packaging"] |
developing_skills | array | ["TypeScript", "RAG pipelines"] |
dealbreaker_detail | array | [{dealbreaker: "on-call", hardness: "absolute"}, {dealbreaker: "relocation", hardness: "strong_preference"}] |
rejection_patterns | array | ["roles that sounded like automation but were actually IT support"] |
These can be set during initial setup or added later with the update_profile tool.
search_jobs builds a search query from your profile and any keywords you provide. Claude then uses that query with its web search capabilities to find real listings. The tool itself does not search the web.
analyze_fit runs a two-step process. First it parses the job description into structured signal (hard requirements, nice-to-haves, red flags, compensation signals, role type) via the Claude API. Then it fetches your portfolio and auto-selects the best resume variant for the detected role type. Claude sees structured signal before raw marketing copy.
parse_jd is the standalone JD parser. Use it independently to pre-process a job description without running the full fit analysis.
log_application, get_applications, and update_application form a job search CRM. Track status (applied, screening, interview, offer, rejected, ghosted), set follow-up dates, and record outcomes.
get_follow_ups shows applications that need attention — where the follow-up date has passed and you're still waiting.
get_application_patterns analyzes your full application history (10+ needed) to find which role types get responses, which skills resonate, and recommends search adjustments.
email_ranked_jobs.py is a standalone Conductor worker script that emails the latest ranked job report. See Email Worker Setup below for configuration.
save_scouted_job saves a job listing found during scouting. It validates the URL to reject search result pages and deduplicates against existing scouted jobs and applications.
get_scouted_jobs returns all scouted listings, optionally filtered to only unranked ones. mark_jobs_ranked marks all unranked jobs as ranked after a ranking report is generated.
add_resume and list_resumes manage multiple resume variants. Each variant targets specific role types (e.g. "AI Engineering" targets ai_engineering and ml_engineering). During fit analysis, the best variant is auto-selected based on the JD's detected role type.
update_profile merges new or changed fields into your existing profile without re-running setup.
get_portfolio and get_resume can be called independently if you want Claude to review just your repos or just your resume.
For the simplest setup, paste your resume directly — no local files needed:
email_ranked_jobs.py sends the ranked job report via Gmail. It is designed to run as a Conductor worker on a schedule.
Credentials live in a gitignored config file — never in the source. Copy the template and fill it in:
Then edit data/email_config.json:
| Key | Required | Default |
|---|---|---|
gmail_user | yes | — |
gmail_app_password | yes | — |
email_to | no | falls back to gmail_user |
email_cap | no | 15 |
The whole data/ folder is gitignored (only .gitkeep and email_config.example.json are tracked), so email_config.json can never be committed by accident.
Why a config file and not environment variables? Depending on how the worker process is launched — notably via Conductor on Windows — environment variables may not be inherited by the child process, causing silent send failures. The config file is read from a path relative to the script, so it works regardless of how the process was spawned. Environment variables (GMAIL_USER, GMAIL_APP_PASSWORD, EMAIL_TO, EMAIL_CAP) are still honored as a fallback when a key is absent from the config file, which is handy for CI or containerized runs. If neither source supplies credentials, the script exits with code 2 and prints setup instructions instead of failing silently.
To generate a Gmail App Password: Google Account → Security → 2-Step Verification → App Passwords. Create one for "Mail". This is not your normal account password. If you ever paste one into a source file by mistake, revoke it from that same screen.
The script reads data/ranked_jobs.md, caps the email to the top 15 ranked listings, sends it, and then deletes ranked_jobs.md so stale rankings are not recycled on the next run.